Embedding-Based Federated Learning with Runtime Governance for Iron Deficiency Prediction

# Embedding-Based Federated Learning with Runtime Governance for Iron Deficiency Prediction††thanks: *These authors contributed equally to this work. ††thanks: For the purpose of open access, the author has applied a Creative Commons Attribution (CC BY) license to any Author Accepted Manuscript version arising.
Source: [https://arxiv.org/html/2605.21563](https://arxiv.org/html/2605.21563)
Simon Deltadahl\*Majid Lotfian DeloueeDaniel KreuterJoseph TaylorAllerdien VisserBloodCounts\! ConsortiumJames H\. F\. RuddNicholas S\. GleadallSuthesh SivapalaratnamFolkert AsselbergsMartijn C\. SchutMichael Roberts

###### Abstract

Recent reviews find that the vast majority of published healthcare federated learning \(FL\) studies never reach real\-world deployment\. We developed an embedding\-based FL pipeline for iron deficiency prediction from routine full blood count \(FBC\) data and deployed it across real institutional environments at Amsterdam University Medical Centre \(AUMC\) and NHS Blood and Transplant \(NHSBT\), two clinical environments that differ markedly in iron deficiency prevalence, ferritin distribution, and subject populations\. A frozen domain\-specific haematology foundation model, DeepCBC, performs site\-local representation extraction, restricting federated training to a compact downstream classifier and substantially reducing recurrent communication relative to full\-encoder federation\. The two clinical datasets are structurally not independent and identically distributed \(non\-IID\), with heterogeneity arising from distinct population differences rather than sampling artefacts\. Runtime governance is enforced by FLA3, a healthcare\-oriented FL platform providing study\-scoped execution, policy\-based authorisation, and signed audit logging\. Standard sample\-size\-weighted aggregation \(FedAvg\) reduced the area under the receiver operating characteristic curve \(ROC\-AUC\) at both sites relative to local\-only training, as the global update was biased towards the larger AUMC distribution\. FedMAP, a personalised aggregation method, raised ROC\-AUC from 0\.9470 to 0\.9594 at AUMC and from 0\.8558 to 0\.8671 at NHSBT relative to local\-only training, achieving the highest macro ROC\-AUC of 0\.9133 and the best macro balanced accuracy overall\. These results support personalised aggregation in clinical federations where client sample size and task relevance diverge substantially\.

## IIntroduction

Full blood count \(FBC\) testing is among the most frequently ordered investigations in clinical practice\. FBC panels do not measure iron stores directly, but indices such as haemoglobin concentration \(HGB\), mean corpuscular volume \(MCV\), mean corpuscular haemoglobin \(MCH\), and red cell distribution width \(RDW\) carry information relevant to iron deficiency\. Prior studies show that machine learning models built from routine laboratory data can predict low ferritin \(a proxy for low body iron stores, and the principal test used for iron deficiency diagnosis\) with useful discriminatory power\[[15](https://arxiv.org/html/2605.21563#bib.bib11),[12](https://arxiv.org/html/2605.21563#bib.bib12),[6](https://arxiv.org/html/2605.21563#bib.bib10),[5](https://arxiv.org/html/2605.21563#bib.bib18)\]\. The central obstacle is not model design alone\. Access to sufficiently broad data is equally limiting: a single institution may capture too narrow a clinical population\[[13](https://arxiv.org/html/2605.21563#bib.bib4)\], and centralisation is constrained by privacy law, institutional governance, and operational trust boundaries\.

Federated learning \(FL\) addresses part of this problem by training models across institutions without pooling raw patient data\[[10](https://arxiv.org/html/2605.21563#bib.bib1),[11](https://arxiv.org/html/2605.21563#bib.bib3),[13](https://arxiv.org/html/2605.21563#bib.bib4)\]\. In healthcare, the challenge extends beyond decentralised optimisation\. Recent review work argues that many published healthcare FL studies remain unsuitable for clinical use citing methodological weaknesses across bias, privacy, generalisation, communication, and governance compliance\[[7](https://arxiv.org/html/2605.21563#bib.bib13),[19](https://arxiv.org/html/2605.21563#bib.bib15),[21](https://arxiv.org/html/2605.21563#bib.bib9)\]\. A complementary systematic review found that only 5\.2% of healthcare FL studies reported real\-life application\[[14](https://arxiv.org/html/2605.21563#bib.bib14)\]\. Genuine clinical deployment remains rare, which motivates empirical reports from operational settings\. Sites also differ in population, indication for testing, laboratory workflow, governance requirements, and network posture, so the problem of not independently and identically distributed data \(non\-IID\) is often structural rather than incidental\.

We report a cross\-institutional study across two BloodCounts\! consortium sites, AUMC and NHSBT\. The system uses the pre\-trained domain\-specific FBC foundation model DeepCBC\[[5](https://arxiv.org/html/2605.21563#bib.bib18)\]for site\-local representation extraction and restricts federated training to a compact downstream classifier\. We compare local\-only training with FedAvg\[[10](https://arxiv.org/html/2605.21563#bib.bib1)\], FedProx\[[8](https://arxiv.org/html/2605.21563#bib.bib2)\], and FedMAP\[[20](https://arxiv.org/html/2605.21563#bib.bib22)\], a personalised aggregation method developed for heterogeneous healthcare federations\. The deployment runs on FLA3\[[21](https://arxiv.org/html/2605.21563#bib.bib9)\], which provides study\-scoped execution, runtime policy enforcement, and auditable logging\.

This paper contributes:

1. 1\.A practical two\-stage design for healthcare FL in which a frozen haematology foundation model handles site\-local representation extraction and only a compact downstream classifier is trained federatively\.
2. 2\.A characterisation of a strongly heterogeneous clinical federation in which sites differ substantially in prevalence, ferritin distribution, and effective positive\-class volume, arising from distinct clinical workflows\.
3. 3\.An empirical demonstration that sample\-size\-weighted aggregation degrades performance at both sites when prevalence and clinical purpose diverge, and that personalised aggregation recovers and exceeds local\-only ROC\-AUC at both sites in this two\-site deployment\.
4. 4\.An operational deployment demonstrating how runtime governance controls, including study scoping, policy\-based authorisation, and signed audit logging, are integrated into a healthcare FL system\.

## IIRelated Work

Federated learning has been applied to medical imaging, electronic health records, and multi\-centre clinical prediction, with recurrent emphasis on data heterogeneity, generalisation, and deployability\[[11](https://arxiv.org/html/2605.21563#bib.bib3),[13](https://arxiv.org/html/2605.21563#bib.bib4),[16](https://arxiv.org/html/2605.21563#bib.bib5),[17](https://arxiv.org/html/2605.21563#bib.bib6)\]\. FedProx introduced a proximal term to stabilise local optimisation under heterogeneous client distributions\[[8](https://arxiv.org/html/2605.21563#bib.bib2)\]\. FedMAP extended this line of work through personalised aggregation that accounts for client relevance under heterogeneous clinical data rather than relying on dataset size alone\[[20](https://arxiv.org/html/2605.21563#bib.bib22)\]\.

Representation transfer is a complementary design consideration\. A domain\-specific model can be pre\-trained once, after which only a lightweight downstream head is adapted, reducing communication volume and simplifying deployment in restrictive institutional environments\[[9](https://arxiv.org/html/2605.21563#bib.bib16)\]\. The present work instantiates this pattern in haematology, coupling a pre\-trained FBC representation model with a federated runtime that enforces study scope, authorisation, and auditability\.

## IIIDatasets and Clinical Heterogeneity

We study two clinically distinct cohorts drawn from BloodCounts\! consortium sites\.

#### AUMC cohort\.

AUMC contributes a hospital\-based cohort in which ferritin testing is performed for the diagnosis of iron deficiency, monitoring or treatment response evaluation, exclusion of overload, or broader inpatient and outpatient workup\. The cohort was56\.3%/56\.3\\text\{\\,\}\\mathrm\{\\char 37\\relax\}\\text\{/\}male and43\.7%/43\.7\\text\{\\,\}\\mathrm\{\\char 37\\relax\}\\text\{/\}female, with median ages of63\.0years63\.0\\text\{\\,\}\\mathrm\{y\}\\mathrm\{e\}\\mathrm\{a\}\\mathrm\{r\}\\mathrm\{s\}and58\.0years58\.0\\text\{\\,\}\\mathrm\{y\}\\mathrm\{e\}\\mathrm\{a\}\\mathrm\{r\}\\mathrm\{s\}, respectively\. The median white blood cell count was6\.71×109cells/l6\.71\\text\{\\times\}\{10\}^\{9\}\\text\{\\,\}\\mathrm\{cells\}\\text\{/\}\\mathrm\{l\}, with74\.8%/74\.8\\text\{\\,\}\\mathrm\{\\char 37\\relax\}\\text\{/\}of records below10×109cells/l10\\text\{\\times\}\{10\}^\{9\}\\text\{\\,\}\\mathrm\{cells\}\\text\{/\}\\mathrm\{l\}\. A reactive testing approach biases towards more severe cases\. Iron deficiency prevalence is lower and replete ferritin values are comparatively high, reflecting higher levels of inflammation within the cohort\. The site is informative for evaluating specificity and generalisation in a clinically complex setting, but it contributes relatively few positive cases\.

#### NHSBT cohort\.

We use data from the INTERVAL randomised controlled trial\[[2](https://arxiv.org/html/2605.21563#bib.bib19)\]assessing the safety of different blood donation frequencies \(8, 10, 12 weeks for males; 12, 14, 16 weeks for females\) over a 24\-month period, with a subgroup also monitored up to 48 months\[[4](https://arxiv.org/html/2605.21563#bib.bib20)\]\. The cohort is approximately balanced by sex \(49\.7%/49\.7\\text\{\\,\}\\mathrm\{\\char 37\\relax\}\\text\{/\}male, median age 46\.2 years;50\.3%/50\.3\\text\{\\,\}\\mathrm\{\\char 37\\relax\}\\text\{/\}female, median age 40\.8 years\)\. This cohort is representative of a healthy population, where iron deficiency is the primary cause of anaemia\. Unlike AUMC’s hospital population, the inflammation burden in this population was low, with a median white blood cell count \(WBC\) of6\.29×109cells/l6\.29\\text\{\\times\}\{10\}^\{9\}\\text\{\\,\}\\mathrm\{cells\}\\text\{/\}\\mathrm\{l\}\(95%/95\\text\{\\,\}\\mathrm\{\\char 37\\relax\}\\text\{/\}of participants with WBC under10×109cells/l10\\text\{\\times\}\{10\}^\{9\}\\text\{\\,\}\\mathrm\{cells\}\\text\{/\}\\mathrm\{l\}\)\. Being a blood donor population, the iron deficiency prevalence was high compared to the general population at around19%/19\\text\{\\,\}\\mathrm\{\\char 37\\relax\}\\text\{/\}\. A universal ferritin testing approach and exclusion of cases with established anaemia biases this cohort towards subclinical cases of iron deficiency, which are associated with milder changes in FBC indices and are thus difficult to detect\.

TABLE I:Dataset statistics per site and split\. Pos\.% is the fraction of samples with ferritin<15µ​g/l<$15\\text\{\\,\}\\mathrm\{\\SIUnitSymbolMicro g\}\\text\{/\}\\mathrm\{l\}$\. Replete ferritin is reported for the training split only\.Table[I](https://arxiv.org/html/2605.21563#S3.T1)summarises the split statistics\. Differences in prevalence, ferritin distribution, and effective positive\-class volume reflect genuine biological and clinical workflow differences between the two populations\. Figs\.[1](https://arxiv.org/html/2605.21563#S3.F1)and[2](https://arxiv.org/html/2605.21563#S3.F2)illustrate the prevalence gap and the marked divergence in iron\-replete ferritin values\. Figure[3](https://arxiv.org/html/2605.21563#S3.F3)shows, for each site, the five embedding dimensions with the largest absolute difference in mean activation between iron\-deficient and iron\-replete samples\. The ranked lists share no overlap\. This divergence likely reflects the differing biological character of the two populations: AUMC’s hospital cohort carries a higher inflammation burden, which shifts FBC indices differently than the low\-inflammation donor population at NHSBT, resulting in distinct discriminative directions in the embedding space\.

#### Label and feature heterogeneity\.

Prevalence differs by nearly an order of magnitude in the training split, from 19\.5% at NHSBT to 2\.8% at AUMC, and the gap persists across validation and test data\. Feature\-level heterogeneity is also evident in the embedding space\. Together, these observations support treating the federation as structurally non\-IID, with heterogeneity driven by clinical workflow and population differences rather than ordinary sampling noise\.

![Refer to caption](https://arxiv.org/html/2605.21563v1/x1.png)Figure 1:Cohort composition at AUMC and NHSBT\. \(a\) Total and iron\-deficient sample counts by split, shown on a log scale\. AUMC contributes larger total sample volume, whereas NHSBT contributes a higher relative burden of iron\-deficient cases\. \(b\) Iron deficiency prevalence by split, showing a persistent prevalence gap between AUMC and NHSBT across training, validation, and test sets\.![Refer to caption](https://arxiv.org/html/2605.21563v1/x2.png)Figure 2:Training\-set ferritin distributions at AUMC and NHSBT for iron\-deficient and iron\-replete groups, shown as median with IQR on a log scale\. The replete ferritin distribution differs markedly across sites, with substantially higher values at AUMC\.![Refer to caption](https://arxiv.org/html/2605.21563v1/x3.png)Figure 3:Top five discriminative embedding dimensions at AUMC and NHSBT, ranked by the absolute difference in class\-conditional mean activation between iron\-deficient and iron\-replete samples\.

## IVMethods

### IV\-ATask Formulation

Each sitekkholds a private dataset𝒟k=\{\(𝐱i\(k\),yi\(k\)\)\}i=1Nk\\mathcal\{D\}\_\{k\}=\\\{\(\\mathbf\{x\}\_\{i\}^\{\(k\)\},y\_\{i\}^\{\(k\)\}\)\\\}\_\{i=1\}^\{N\_\{k\}\}, where𝐱i\(k\)\\mathbf\{x\}\_\{i\}^\{\(k\)\}is the patient representation derived from local FBC data and

yi\(k\)=𝟏​\[ferritini\(k\)<15µ​g/l\]y\_\{i\}^\{\(k\)\}=\\mathbf\{1\}\[\\mathrm\{ferritin\}\_\{i\}^\{\(k\)\}<$15\\text\{\\,\}\\mathrm\{\\SIUnitSymbolMicro g\}\\text\{/\}\\mathrm\{l\}$\]indicates iron deficiency in line with world health organisation guidelines\[[18](https://arxiv.org/html/2605.21563#bib.bib17)\]\. The objective is to learn a classifierf𝜽:ℝd→\[0,1\]f\_\{\\bm\{\\theta\}\}:\\mathbb\{R\}^\{d\}\\to\[0,1\]without transferring raw patient records between institutions, where the embedding dimensionalityddis determined by the site\-local feature extractor \(Section[IV\-B](https://arxiv.org/html/2605.21563#S4.SS2)\)\.

### IV\-BEmbedding Extraction

The first stage uses DeepCBC, a foundation model trained previously on large\-scale flow cytometry and impedance data underlying the FBC \(rawFBC data\)\[[5](https://arxiv.org/html/2605.21563#bib.bib18)\]\. At deployment time only the frozen encoder is used\. Given a standardised local FBC input𝐱i\\mathbf\{x\}\_\{i\}, the encoder defines a latent posterior

qϕ​\(𝐳∣𝐱i\)=𝒩​\(𝝁ϕ​\(𝐱i\),diag​\(𝝈ϕ2​\(𝐱i\)\)\)\.q\_\{\\phi\}\(\\mathbf\{z\}\\mid\\mathbf\{x\}\_\{i\}\)=\\mathcal\{N\}\\\!\\big\(\\bm\{\\mu\}\_\{\\phi\}\(\\mathbf\{x\}\_\{i\}\),\\mathrm\{diag\}\(\\bm\{\\sigma\}^\{2\}\_\{\\phi\}\(\\mathbf\{x\}\_\{i\}\)\)\\big\)\.For downstream classification we use the posterior mean as a deterministic embedding,

𝐳i=𝝁ϕ​\(𝐱i\)∈ℝ256\.\\mathbf\{z\}\_\{i\}=\\bm\{\\mu\}\_\{\\phi\}\(\\mathbf\{x\}\_\{i\}\)\\in\\mathbb\{R\}^\{256\}\.This choice removes stochasticity at inference time and provides a fixed\-length representation for each patient\. Raw FBC inputs and patient\-level embeddings remain inside the institutional environment, the foundation model is distributed once as a local artefact, and recurrent federated communication is limited to classifier parameters\.

This two\-stage decomposition is motivated by both optimisation and deployment constraints\. Adaptation is concentrated at the downstream decision boundary rather than requiring joint representation learning across highly heterogeneous cohorts\. Communication overhead is reduced relative to a federation that repeatedly synchronises a larger encoder\. In this deployment, recurrent communication is limited to the downstream multilayer perceptron \(MLP\) classifier parameters, whereas synchronising the full DeepCBC encoder at the same cadence would require transmitting roughly three orders of magnitude more parameters per round\.

### IV\-CFederated Classifier and Aggregation

The downstream classifier is a two\-hidden\-layer MLP \(256→128→64→1256\\rightarrow 128\\rightarrow 64\\rightarrow 1\) with batch normalisation, ReLU activations, and dropout \(p=0\.3p=0\.3\), totalling 41,601 trainable parameters\. Local\-only training serves as the non\-federated baseline at each site\.

Three federated aggregation strategies are compared:

- •FedAvg: the standard sample\-size\-weighted averaging baseline\[[10](https://arxiv.org/html/2605.21563#bib.bib1)\]\.
- •FedProx: local optimisation with a proximal penaltyμ2​‖𝜽−𝜽\(t\)‖22\\frac\{\\mu\}\{2\}\\\|\\bm\{\\theta\}\-\\bm\{\\theta\}^\{\(t\)\}\\\|\_\{2\}^\{2\}to reduce client drift under heterogeneity\[[8](https://arxiv.org/html/2605.21563#bib.bib2)\]\.
- •FedMAP: a personalised aggregation framework in which local optimisation is cast as maximum a posteriori \(MAP\) estimation under a learned regulariserR​\(θ;μg,ψ\)R\(\\theta;\\mu\_\{g\},\\psi\), and server aggregation uses posterior\-informed weights rather than sample count alone\[[20](https://arxiv.org/html/2605.21563#bib.bib22)\]\.

FedMAP is summarised here; the full derivation is given in the original paper\. FedMAP learns a global regulariser of the form

R​\(θ;θg,ψ\)=fψ​\(θ,θg\)\+α​∥θ−θg∥2\+ϵ​\(∥θ∥2\+∥θg∥2\),R\(\\theta;\\theta\_\{g\},\\psi\)=f\_\{\\psi\}\(\\theta,\\theta\_\{g\}\)\+\\alpha\\lVert\\theta\-\\theta\_\{g\}\\rVert^\{2\}\+\\epsilon\(\\lVert\\theta\\rVert^\{2\}\+\\lVert\\theta\_\{g\}\\rVert^\{2\}\),wherefψf\_\{\\psi\}is an input\-convex neural network \(ICNN\),θg\\theta\_\{g\}denotes the global model parameters, andψ\\psiparameterises the learned regularisation landscape\. At each round, sitekkupdates its local classifier by MAP optimisation and returns both parameters and an aggregation weight, presented in summary form from:

ωk\(t\)∝P​\(𝒟k∣θk\(t\+1\)\)​exp⁡\(−R​\(θk\(t\+1\);θg\(t\),ψ\(t\)\)\),\\omega\_\{k\}^\{\(t\)\}\\propto P\(\\mathcal\{D\}\_\{k\}\\mid\\theta\_\{k\}^\{\(t\+1\)\}\)\\exp\\\!\\big\(\-R\(\\theta\_\{k\}^\{\(t\+1\)\};\\theta\_\{g\}^\{\(t\)\},\\psi^\{\(t\)\}\)\\big\),where𝒟k\\mathcal\{D\}\_\{k\}is the local dataset at sitekk\. This weight favours local models that fit their data well and remain close to the globally learned prior\. The server then updates the global model by weighted averaging and refinesψ\\psithrough gradient steps on the regulariser objective\.

All federated methods used 50 communication rounds, batch size 256, learning rate10−310^\{\-3\}, 10 local epochs per round, and early stopping with patience 3\. Thresholds were tuned on the validation split \(Table[I](https://arxiv.org/html/2605.21563#S3.T1)\)\. FedProx used proximal coefficientμp=0\.05\\mu\_\{p\}=0\.05\. FedMAP used an ICNN learning rate of10−510^\{\-5\}and 3 ICNN steps per round\.

## VFLA3Platform and Runtime Governance

Governance in this deployment is enforced at runtime rather than recorded as a static agreement outside the system\. The federation therefore runs on FLA3\[[21](https://arxiv.org/html/2605.21563#bib.bib9)\], a healthcare\-oriented federated learning platform built around Flower\-based\[[1](https://arxiv.org/html/2605.21563#bib.bib21)\]orchestration with explicit authorisation and auditing\.

Each study is scoped as a distinct federation with explicit participant roles\. The orchestration layer separates central coordination from site\-local execution, corresponding broadly to Flower SuperLink and SuperNode responsibilities in the platform design\. Authorisation is policy\-driven and compliant with the eXtensible Access Control Markup Language \(XACML\), with deny\-by\-default semantics: an action is permitted only if an explicit rule grants it\. Key execution events are recorded in an append\-only signed audit trail, allowing institutions to reconstruct who requested a training action, under which policy, and at what time\. In this deployment, a site\-local node attempting to push parameters outside its assigned aggregation window is denied by the XACML policy engine before any network transmission occurs; the denial is recorded in the audit log with the policy rule identifier and timestamp\.

Participating organisations in multi\-institutional healthcare FL often require assurance beyond local data retention\. Role\-bound execution, temporal scoping, policy compliance, and post hoc accountability are distinct requirements that governance documentation alone cannot satisfy\. FLA3provides those controls as constituents of the training runtime rather than as an external wrapper\.

## VIExperiments, Results and Discussion

Each site maintained separate train, validation, and test splits\. The primary metric was ROC–AUC, reported separately for AUMC and NHSBT and summarised as a macro average across sites\. Balanced accuracy was also reported as class distributions differed sharply across sites\. Threshold\-dependent metrics used thresholds selected on the validation split\. Macro ROC–AUC and macro balanced accuracy were computed as the mean of the two per\-site values\. Confidence intervals are reported to summarise uncertainty\.

The primary interpretive question concerns whether federation improves performance at either site without material harm to the other, and whether personalised aggregation reduces the AUMC–NHSBT performance gap\.

TABLE II:Test\-set performance by site\. FedMAP achieves the highest ROC–AUC at both sites and the highest macro balanced accuracy overall\. CI: confidence interval\.Table[II](https://arxiv.org/html/2605.21563#S6.T2)reports the main test\-set results\. FedAvg reduced ROC–AUC relative to local\-only training at both AUMC and NHSBT\. AUMC contributes roughly 1\.4 times the total training samples of NHSBT, so sample\-size\-weighted averaging biases the global update towards the AUMC distribution\. Given the marked differences in prevalence and embedding\-level class structure between the two cohorts, this bias degrades the decision boundary at both sites rather than improving either\. FedProx recovered part of this loss but did not achieve the best overall performance; its proximal penalty stabilises local optimisation around a shared global model, but aggregation remains insensitive to client task relevance\. It therefore addresses client drift without resolving the underlying mismatch between sample\-size weighting and structural clinical heterogeneity\. FedMAP achieved the highest ROC–AUC at both sites, improving from 0\.9470 to 0\.9594 at AUMC and from 0\.8558 to 0\.8671 at NHSBT compared to local\-only training, with the highest macro ROC–AUC of 0\.9133\. The discrepancy in performance in terms of ROC\-AUC is likely to reflect difference in iron deficiency case severity between sites, resulting from the different ferritin testing approach \(universal vs reactive\) described above\.

Balanced accuracy showed a similar but not identical pattern\. At NHSBT, all federated methods improved on the local\-only baseline, with FedMAP reaching 0\.7879\. At AUMC, local\-only training remained highest at 0\.8964; FedMAP was close at 0\.8899 and outperformed FedAvg and FedProx\. The small balanced accuracy gap at AUMC \(0\.0065\) likely reflects threshold sensitivity under the site’s low positive prevalence: personalised aggregation shifts the global prior in a direction that improves discrimination \(ROC–AUC\) without a corresponding shift in the optimal classification threshold, so threshold\-dependent metrics do not fully capture the gain\. FedMAP achieved the highest macro balanced accuracy of 0\.8389\. Of note, in these experiments we do not allow for threshold re\-optimisation post\-aggregation by site and use the World Health Organisation recommended threshold of15µ​g/l15\\text\{\\,\}\\mathrm\{\\SIUnitSymbolMicro g\}\\text\{/\}\\mathrm\{l\}; however, due to the higher levels of inflammation in the AUMC hospital cohort we expect an iron\-status independent right shift in ferritin values and a corresponding loss of test sensitivity\. This effect is demonstrated in the higher median ferritin in iron repletion seen in the AUMC cohort \(Figure[2](https://arxiv.org/html/2605.21563#S3.F2)\)\. In clinical practice a higher threshold of30µ​g/l30\\text\{\\,\}\\mathrm\{\\SIUnitSymbolMicro g\}\\text\{/\}\\mathrm\{l\}is frequently employed as a crude form of threshold re\-optimisation to adjust for this problem\[[3](https://arxiv.org/html/2605.21563#bib.bib23)\]\.

These findings support the view that personalised aggregation is beneficial when client sample size and task relevance diverge across sites\.

## VIILessons Learned and Practical Implications

#### Lesson 1: structural heterogeneity should be treated as a clinical property, not only a statistical nuisance

We applied the same ferritin threshold to define iron deficiency for NHSBT and AUMC, but their cohorts arise from different clinical workflows: NHSBT reflects donor screening, whereas AUMC reflects hospital\-based testing and monitoring\. The resulting prevalence and ferritin\-distribution differences indicate that the sites do not represent the same task distribution\. For healthcare FL studies, reporting class balance alone is insufficient; the clinical origin of the label must also be described\.

#### Lesson 2: embedding\-based federation offers a practical deployment pattern for hospitals

A frozen domain\-specific foundation model can be used for local representation extraction, reducing the federated task to a compact downstream classifier\. This lowers communication burden and simplifies deployment inside institutional environments\. It retains the benefits of domain\-specific pretraining without repeated synchronisation of a larger encoder\. In this deployment, recurrent communication is limited to the downstream MLP parameters, whereas synchronising the full DeepCBC encoder at the same cadence would require transmitting roughly three orders of magnitude more parameters per round\. A further practical advantage is data efficiency: individual clinical sites may hold insufficient data to train a large representation model from scratch, whereas a frozen pretrained encoder transfers domain knowledge that would otherwise require far larger local datasets to acquire\. In hospital settings where both infrastructure constraints and limited local data volumes make full\-model federation difficult, separating representation learning from federated task adaptation is a practical design choice\.

#### Lesson 3: personalised aggregation may help when client size and client relevance diverge

Larger local sample volume does not necessarily imply greater relevance to the shared task\. The two cohorts differ in clinical purpose, case mix, and representation\-level class structure, and FedMAP outperformed size\-weighted baselines across the reported metrics\. Healthcare FL evaluations should consider whether aggregation reflects task alignment rather than assuming that larger sites should dominate the global update\.

#### Lesson 4: governance should be operationalised at runtime

Multi\-institutional healthcare federation requires more than local data retention\. Without runtime enforcement, a site\-local node can transmit parameters outside its assigned aggregation window, exfiltrate embeddings, or participate beyond its authorised study scope, flaws that static data\-sharing agreements cannot prevent after the fact\. In this deployment, such violations are denied by the XACML policy engine before any network transmission occurs and recorded in the audit log\. This runtime enforcement was a practical prerequisite for executing the study across international borders: AUMC and NHSBT operate under distinct regulatory frameworks \(Dutch and UK clinical governance respectively\), and neither institution could have participated without verifiable, post hoc accountable controls over how their data and model parameters were used\. Governance documentation alone cannot provide that assurance; it must be built into the execution layer\.

#### Limitations

Privacy attack analyses, including membership inference and gradient inversion, were not conducted in this deployment phase and constitute an important direction for future work\. The embedding\-only communication pattern reduces the gradient inversion attack surface relative to raw\-feature federation, but does not eliminate privacy risk entirely\. Runtime overhead for the personalised aggregation procedure was not quantified; future work will profile wall\-clock training time per round, including a comparison against the cost of synchronising the full DeepCBC encoder to characterise the communication and compute savings of the embedding\-based design\. The embedding\-space heterogeneity comparison is qualitative, and formal paired significance testing between methods was outside the scope\. As evidence derives from a single two\-site deployment, broader validation across larger consortium sites would strengthen the generalisability of these findings\. Calibration was not evaluated in this deployment phase\. This is a limitation because clinical prediction models require reliable probability estimates in addition to discrimination; future deployment should include site\-specific calibration assessment and recalibration where necessary before clinical use\. Because DeepCBC is frozen, a maintained deployment would require monitoring of embedding distributions for temporal drift in laboratory practice, analyser hardware, preprocessing, or patient case mix\. Future work should examine whether learned FedMAP aggregation weights can serve as quantitative indicators of inter\-site heterogeneity\.

## VIIIEthics and Responsible Research

This study uses clinical laboratory data in a privacy\-constrained setting\. Raw patient data were not centralised for model training; raw FBC measurements and patient\-level embeddings remained within each institution\. Only downstream classifier parameters were exchanged during federation\. The INTERVAL trial is compiled into the Blood Donors Studies BioResource \(BDSB\) with Research Ethics Committee \(REC\) reference 20/EE/0115\. The Medical Ethics Review Committee of Amsterdam University Medical Centres \(registered with the US Office for Human Research Protections as IRB00013752; FWA number FWA00032965\) reviewed the study protocol and determined that the Medical Research Involving Human Subjects Act \(WMO\) does not apply to this research; therefore, formal ethical approval and informed consent were waived\.

## IXConclusion

We present an operational FL deployment for iron deficiency prediction from FBC foundation\-model embeddings across AUMC and NHSBT\. The study brings together three elements that are often treated separately in the literature: domain\-specific representation transfer, structural healthcare heterogeneity, and runtime governance controls\. Standard sample\-size\-weighted federation did not improve discrimination across sites\. FedMAP, a personalised FL algorithm, achieved the strongest overall performance, with the highest ROC–AUC at both AUMC and NHSBT and the best macro balanced accuracy\. These findings, drawn from a two\-site deployment, support the case for personalised aggregation in operational clinical federations with substantial structural heterogeneity\.

## Acknowledgment

F\. Zhang, S\. Deltadahl, D\. Kreuter, J\. Taylor, N\. S\. Gleadall, S\. Sivapalaratnam, and M\. Roberts have received support from the Trinity Challenge grant awarded to establish the BloodCounts\! consortium, along with NIHR UCLH Biomedical Research Centre, the NIHR Cambridge Biomedical Research Centre, National Health Service Blood and Transplant \(NHSBT\) and the Medical Research Council\. D\. Kreuter and J\. Taylor receive support from MRC GAP Fund \(UKRI/814\)\. N\. S\. Gleadall has been supported by NHSBT grant 1701\-GEN\. M\. Roberts is additionally supported by the British Heart Foundation \(TA/F/20/210001\)\. M\. Lotfian Delouee acknowledges support from the LabGPT project, funded by Amsterdam UMC Innovation Funding\.

## Conflicts of Interest

The authors declare no conflicts of interest relevant to this work\.

## Code Availability

## BloodCounts\! Consortium Members

Martijn Schut1, Folkert Asselbergs1, Sujoy Kar2, Suthesh Sivapalaratnam3, Sophie Williams3, Mickey Koh4, Yvonne Henskens5, Norbert C\.J\. de Wit5, Umberto D’Alessandro6, Bubacarr Bah6, Ousman Secka6, Parashkev Nachev7, Rajeev Gupta7, Sara Trompeter7, Nancy Boeckx8, Christine van Laer8, Gordon A\. Awandare9, Kwabena Sarpong9, Lucas Amenga\-Etego9, Mathie Leers10, Mirelle Huijskens10, Samuel McDermott11, Willem H\. Ouwehand12, James Rudd13, Carola\-Bibiane Schönlieb11, Nicholas Gleadall12,14,15, and Michael Roberts11,13\.

1Amsterdam University Medical Centre, Amsterdam, Netherlands\. 2Apollo Hospitals, Chennai, India\. 3Barts Health NHS Trust, London, United Kingdom\. 4Health Services Authority, Singapore\. 5Maastricht University Medical Centre, Maastricht, Netherlands\. 6MRC The Gambia Unit, Banjul, The Gambia\. 7University College London Hospitals, London, United Kingdom\. 8University Hospitals Leuven, Leuven, Belgium\. 9West African Centre for Cell Biology of Infectious Pathogens, Accra, Ghana\. 10Zuyderland Medical Center, Zuyderland, Netherlands\. 11Department of Applied Mathematics and Theoretical Physics, University of Cambridge, UK\. 12NHS Blood and Transplant, Cambridge, UK\. 13Department of Medicine, University of Cambridge, UK\. 14Victor Phillip Dahdaleh Heart and Lung Research Institute, University of Cambridge, UK\. 15Department of Haematology, University of Cambridge, UK\.

## References

- \[1\]D\. J\. Beutel, T\. Topal, A\. Mathur, X\. Qiu, J\. Fernandez\-Marques, Y\. Gao, L\. Sani, K\. H\. Li, T\. Parcollet, P\. P\. B\. de Gusmão, and N\. D\. Lane\(2020\)Flower: a friendly federated learning research framework\.External Links:2007\.14390,[Link](https://arxiv.org/abs/2007.14390)Cited by:[§V](https://arxiv.org/html/2605.21563#S5.p1.1)\.
- \[2\]E\. Di Angelantonio, S\. G\. Thompson, S\. Kaptoge, C\. Moore, M\. Walker, J\. Armitage, W\. H\. Ouwehand, D\. J\. Roberts, J\. Danesh, E\. Di Angelantonio, S\. G\. Thompson, S\. Kaptoge, C\. Moore, M\. Walker, J\. Armitage, W\. H\. Ouwehand, D\. J\. Roberts, J\. Danesh, J\. Armitage, J\. Danesh, E\. Di Angelantonio, J\. Donovan, I\. Ford, R\. Henry, B\. J\. Hunt, B\. Le Huray, S\. Mehenny, G\. Miflin, C\. Moore, W\. H\. Ouwehand, J\. Green, D\. J\. Roberts, M\. Stredder, S\. G\. Thompson, M\. Walker, N\. A\. Watkins, A\. McDermott, C\. Ronaldson, C\. Thomson, Z\. Tolkien, L\. Williamson, D\. Allen, J\. Danesh, E\. Di Angelantonio, R\. Henry, S\. Mehenny, C\. Moore, W\. H\. Ouwehand, D\. J\. Roberts, J\. Sambrook, M\. Walker, T\. Hammerton, C\. Thomson, Z\. Tolkien, D\. Allen, D\. Bruce, F\. Choudry, E\. Di Angelantonio, C\. Ghevaert, K\. Johnston, A\. Kelly, A\. King, S\. Mehenny, G\. Miflin, A\. Mo, C\. Moore, W\. H\. Ouwehand, L\. Page, P\. Richardson, D\. J\. Roberts, J\. Sambrook, P\. Senior, Y\. Umrania, M\. Walker, H\. Wong, S\. Kaptoge, G\. Murphy, A\. C\. Newland, K\. Wheatley, M\. Greaves, M\. Turner, T\. Aziz, R\. Brain, C\. Davies, R\. Turner, P\. Wakeman, A\. Dent, A\. Wakeman, B\. Anthony, D\. Bland, W\. Parrondo, H\. Vincent, C\. Weatherill, A\. Forsyth, C\. Butterfield, T\. Wright, K\. Ellis, K\. Johnston, P\. Poynton, C\. Brooks, E\. Martin, L\. Littler, L\. Williams, D\. Blair, K\. Ackerley, L\. Woods, S\. Stanley, G\. Walsh, G\. Franklin, C\. Howath, S\. Sharpe, D\. Smith, L\. Botham, C\. Williams, C\. Alexander, G\. Sowerbutts, D\. Furnival, M\. Thake, S\. Patel, C\. Roost, S\. Sowerby, M\. J\. Appleton, E\. Bays, G\. Bowyer, S\. Clarkson, S\. Halson, K\. Holmes, G\. Humphries, K\. Johnston, L\. Parvin\-Cooper, J\. Towler, J\. Addy, P\. Barrass, L\. Stennett, S\. Burton, H\. Dingwall, R\. Henry, V\. Clarke, M\. Potton, C\. Thomson, T\. Bolton, M\. Daynes, S\. Halson, S\. Spackman, M\. Walker, A\. Momodu, J\. Fenton, A\. King, O\. Muhammed, N\. Oates, T\. Peakman, C\. Ryan, K\. Spreckley, C\. Stubbins, J\. Williams, J\. Brennan, C\. Mochon, S\. Taylor, K\. Warren, S\. Kaptoge, S\. G\. Thompson, E\. Di Angelantonio, C\. Moore, J\. Mant, W\. H\. Ouwehand, S\. G\. Thompson, J\. Danesh, and D\. J\. Roberts\(2017\-11\)Efficiency and safety of varying the frequency of whole blood donation \(interval\): a randomised trial of 45 000 donors\.390\(10110\),pp\. 2360–2371\.External Links:[Document](https://dx.doi.org/10.1016/s0140-6736%2817%2931928-1),ISSN 0140\-6736,[Link](http://dx.doi.org/10.1016/S0140-6736(17)31928-1)Cited by:[§III](https://arxiv.org/html/2605.21563#S3.SS0.SSS0.Px2.p1.6)\.
- \[3\]A\. Fletcher, A\. Forbes, N\. Svenson, and D\. W\. Thomas\(2022\)Guideline for the laboratory diagnosis of iron deficiency in adults \(excluding pregnancy\) and children\.196\(3\),pp\. 523–529\.Note:A British Society for Haematology Good Practice PaperExternal Links:[Document](https://dx.doi.org/https%3A//doi.org/10.1111/bjh.17900),[Link](https://onlinelibrary.wiley.com/doi/abs/10.1111/bjh.17900),https://onlinelibrary\.wiley\.com/doi/pdf/10\.1111/bjh\.17900Cited by:[§VI](https://arxiv.org/html/2605.21563#S6.p4.2)\.
- \[4\]S\. Kaptoge, E\. D\. Angelantonio, C\. Moore, M\. Walker, J\. Armitage, W\. H\. Ouwehand, D\. J\. Roberts, J\. Danesh, S\. G\. Thompson, S\. Kaptoge, E\. D\. Angelantonio, C\. Moore, M\. Walker, J\. Armitage, W\. H\. Ouwehand, D\. J\. Roberts, J\. Danesh, S\. G\. Thompson, J\. Donovan, I\. Ford, R\. Henry, B\. J\. Hunt, B\. l\. Huray, S\. Mehenny, G\. Miflin, J\. Green, M\. Stredder, N\. A\. Watkins, A\. McDermott, C\. Ronaldson, C\. Thomson, Z\. Tolkien, L\. Williamson, D\. Allan, J\. Sambrook, T\. Hammerton, D\. Bruce, F\. Choudry, C\. Ghvaert, K\. Jonston, A\. Kelly, A\. King, A\. Mo, L\. Page, P\. Richardson, P\. Senior, Y\. Umrania, H\. Wong, B\. Burchell, J\. Gallacher, G\. Murphy, A\. C\. Newland, K\. Wheatley, M\. Greaves, M\. Turner, T\. Aziz, R\. Brain, C\. Davies, R\. Turner, P\. Wakeman, A\. Dent, A\. Wakeman, B\. Anthony, D\. Bland, W\. H\. Parrondo, H\. Vincent, C\. Weatherill, A\. Forsyth, C\. Butterfield, T\. Wright, K\. Ellis, K\. Johnston, P\. Poynton, C\. Brooks, E\. Martin, L\. Littler, L\. Williamson, D\. Blair, K\. Ackerley, L\. Woods, S\. Stanley, G\. Walsh, G\. Franklin, C\. Howath, S\. Sharpe, D\. Smith, L\. Botham, C\. Williams, C\. Alexander, G\. Sowerbutts, D\. Furnival, M\. Thake, S\. Patel, C\. Roost, S\. Sowerby, M\. J\. Appleton, E\. Bays, G\. Bowyer, S\. Clarkson, S\. Halson, K\. Holmes, G\. Humphreys, L\. Parvin\-Cooper, J\. Towler, J\. Addy, P\. Barrass, L\. Stennett, S\. Burton, H\. Dingwell, V\. Clarke, M\. Potton, T\. Bolton, M\. Daynes, S\. Halson, S\. Spackman, M\. Walker, A\. Momodu, J\. Fenton, A\. King, O\. Muhammad, N\. Oates, T\. Peakman, C\. Ryan, K\. Spreckley, C\. Stubbins, J\. Williams, J\. Brannan, C\. Mochon, S\. Taylor, K\. Warren, S\. Kaptoge, E\. D\. Angelantonio, J\. Mant, W\. H\. Ouwehand, S\. G\. Thompson, J\. Danesh, and D\. J\. Roberts\(2019\-10\)Longer\-term efficiency and safety of increasing the frequency of whole blood donation \(INTERVAL\): extension study of a randomised trial of 20 757 blood donors\.6\(10\),pp\. e510–e520\(English\)\.Note:Publisher: ElsevierExternal Links:ISSN 2352\-3026,[Link](https://www.thelancet.com/journals/lanhae/article/PIIS2352-3026(19)30106-1/fulltext),[Document](https://dx.doi.org/10.1016/S2352-3026%2819%2930106-1)Cited by:[§III](https://arxiv.org/html/2605.21563#S3.SS0.SSS0.Px2.p1.6)\.
- \[5\]D\. Kreuter, J\. Taylor, S\. Deltadahl, M\. Besser, J\. R\. Bradley, J\. Gilbey, S\. Kaptoge, N\. Kingston, W\. H\. Ouwehand, D\. Roberts, O\. Shamardina, K\. E\. Stirrups, D\. Swinkels, D\. Vuckovic, S\. N\. BioResource, S\. Trial, B\. Consortium, J\. H\. Rudd, E\. D\. Angelantonio, C\. Schönlieb, P\. Nachev, S\. Sivapalaratnam, N\. S\. Gleadall, and M\. Roberts\(2025\)Artificial intelligence for pre\-anaemic iron deficiency detection using rich complete blood count data\.External Links:[Document](https://dx.doi.org/10.1101/2025.06.18.25329494),[Link](https://www.medrxiv.org/content/early/2025/06/24/2025.06.18.25329494),https://www\.medrxiv\.org/content/early/2025/06/24/2025\.06\.18\.25329494\.full\.pdfCited by:[§I](https://arxiv.org/html/2605.21563#S1.p1.1),[§I](https://arxiv.org/html/2605.21563#S1.p3.1),[§IV\-B](https://arxiv.org/html/2605.21563#S4.SS2.p1.1)\.
- \[6\]S\. Kurstjens, I\. Belov, W\. de Kort, M\. van de Schootbrugge, W\. Oosterhuis, and J\. A\. van Balveren\(2022\)Automated prediction of low ferritin concentrations using a machine learning algorithm based on routine laboratory test results\.Clinical Chemistry and Laboratory Medicine \(CCLM\)60\(7\),pp\. e173–e176\.External Links:[Document](https://dx.doi.org/10.1515/cclm-2021-1215)Cited by:[§I](https://arxiv.org/html/2605.21563#S1.p1.1)\.
- \[7\]M\. Li, P\. Xu, J\. Hu, Z\. Tang, and G\. Yang\(2025\)From challenges and pitfalls to recommendations and opportunities: implementing federated learning in healthcare\.Medical Image Analysis101,pp\. 103497\.External Links:ISSN 1361\-8415,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/j.media.2025.103497),[Link](https://www.sciencedirect.com/science/article/pii/S1361841525000453)Cited by:[§I](https://arxiv.org/html/2605.21563#S1.p2.1)\.
- \[8\]T\. Li, A\. K\. Sahu, M\. Zaheer, M\. Sanjabi, A\. Talwalkar, and V\. Smith\(2020\)Federated optimization in heterogeneous networks\.InProceedings of Machine Learning and Systems,Vol\.2,pp\. 429–450\.Cited by:[§I](https://arxiv.org/html/2605.21563#S1.p3.1),[§II](https://arxiv.org/html/2605.21563#S2.p1.1),[2nd item](https://arxiv.org/html/2605.21563#S4.I1.i2.p1.1)\.
- \[9\]S\. Malaviya, M\. Shukla, and S\. Lodha\(2023\-22–25 Aug\)Reducing communication overhead in federated learning for pre\-trained language models using parameter\-efficient finetuning\.InProceedings of The 2nd Conference on Lifelong Learning Agents,S\. Chandar, R\. Pascanu, H\. Sedghi, and D\. Precup \(Eds\.\),Proceedings of Machine Learning Research, Vol\.232,pp\. 456–469\.External Links:[Link](https://proceedings.mlr.press/v232/malaviya23a.html)Cited by:[§II](https://arxiv.org/html/2605.21563#S2.p2.1)\.
- \[10\]H\. B\. McMahan, E\. Moore, D\. Ramage, S\. Hampson, and B\. A\. y Arcas\(2017\)Communication\-efficient learning of deep networks from decentralized data\.InProceedings of the 20th International Conference on Artificial Intelligence and Statistics \(AISTATS\),Proceedings of Machine Learning Research, Vol\.54,pp\. 1273–1282\.Cited by:[§I](https://arxiv.org/html/2605.21563#S1.p2.1),[§I](https://arxiv.org/html/2605.21563#S1.p3.1),[1st item](https://arxiv.org/html/2605.21563#S4.I1.i1.p1.1)\.
- \[11\]N\. Rieke, J\. Hancox, W\. Li, F\. Milletari, H\. R\. Roth, S\. Albarqouni, S\. Bakas, M\. N\. Galtier, B\. A\. Landman, K\. Maier\-Hein, S\. Ourselin, M\. Sheller, R\. M\. Summers, A\. Trask, D\. Xu, M\. Baust, and M\. J\. Cardoso\(2020\)The future of digital health with federated learning\.npj Digital Medicine3\(1\),pp\. 119\.External Links:[Document](https://dx.doi.org/10.1038/s41746-020-00323-1)Cited by:[§I](https://arxiv.org/html/2605.21563#S1.p2.1),[§II](https://arxiv.org/html/2605.21563#S2.p1.1)\.
- \[12\]P\. R\. Sarma\(1990\)Red cell indices\.InClinical Methods: The History, Physical, and Laboratory Examinations,H\. K\. Walker, W\. D\. Hall, and J\. W\. Hurst \(Eds\.\),Cited by:[§I](https://arxiv.org/html/2605.21563#S1.p1.1)\.
- \[13\]M\. J\. Sheller, B\. Edwards, G\. A\. Reina, J\. Martin, S\. Pati, A\. Kotrotsou, M\. Milchenko, W\. Xu, D\. Marcus, R\. R\. Colen, and S\. Bakas\(2020\)Federated learning in medicine: facilitating multi\-institutional collaborations without sharing patient data\.Scientific Reports10\(1\),pp\. 12598\.External Links:[Document](https://dx.doi.org/10.1038/s41598-020-69250-1)Cited by:[§I](https://arxiv.org/html/2605.21563#S1.p1.1),[§I](https://arxiv.org/html/2605.21563#S1.p2.1),[§II](https://arxiv.org/html/2605.21563#S2.p1.1)\.
- \[14\]Z\. L\. Teo, L\. Jin, N\. Liu, S\. Li, D\. Miao, X\. Zhang, W\. Y\. Ng, T\. F\. Tan, D\. M\. Lee, K\. J\. Chua, J\. Heng, Y\. Liu, R\. S\. M\. Goh, and D\. S\. W\. Ting\(2024\)Federated machine learning in healthcare: a systematic review on clinical applications and technical architecture\.Cell Reports MedicinePatternsmedRxivThe LancetThe Lancet HaematologyBritish Journal of Haematology5\(2\),pp\. 101419\.External Links:ISSN 2666\-3791,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/j.xcrm.2024.101419),[Link](https://www.sciencedirect.com/science/article/pii/S2666379124000429)Cited by:[§I](https://arxiv.org/html/2605.21563#S1.p2.1)\.
- \[15\]W\. G\. Thompson, T\. Meola, M\. Lipkin, and M\. L\. Freedman\(1988\)Red cell distribution width, mean corpuscular volume, and transferrin saturation in the diagnosis of iron deficiency\.Archives of Internal Medicine148\(10\),pp\. 2128–2130\.External Links:[Document](https://dx.doi.org/10.1001/archinte.1988.00380100090019)Cited by:[§I](https://arxiv.org/html/2605.21563#S1.p1.1)\.
- \[16\]A\. Vaid, S\. K\. Jaladanki, J\. Xu, S\. Teng, A\. Kumar, S\. Lee, S\. Somani, I\. Paranjpe, J\. K\. De Freitas, T\. Wanyan, K\. W\. Johnson, M\. Bicak, E\. Klang, Y\. J\. Kwon, A\. Costa, S\. Zhao, R\. Miotto, A\. W\. Charney, E\. Böttinger, Z\. A\. Fayad, G\. N\. Nadkarni, F\. Wang, and B\. S\. Glicksberg\(2021\)Federated learning of electronic health records to improve mortality prediction in hospitalized patients with covid\-19: machine learning approach\.JMIR Medical Informatics9\(1\),pp\. e24207\.External Links:[Document](https://dx.doi.org/10.2196/24207)Cited by:[§II](https://arxiv.org/html/2605.21563#S2.p1.1)\.
- \[17\]S\. Warnat\-Herresthal, H\. Schultze, K\. L\. Shastry, S\. Manamohan, S\. Mukherjee, V\. Garg, R\. Sarveswara, N\. Feierberg, U\. Hübner, K\. Müller,et al\.\(2021\)Swarm learning for decentralized and confidential clinical machine learning\.Nature594\(7862\),pp\. 265–270\.External Links:[Document](https://dx.doi.org/10.1038/s41586-021-03583-3)Cited by:[§II](https://arxiv.org/html/2605.21563#S2.p1.1)\.
- \[18\]World Health Organisation\(2020\)WHO guideline on use of ferritin concentrations to assess iron status in individuals and populations\.WHO Guidelines Approved by the Guidelines Review Committee,World Health Organization,Geneva\.External Links:[Link](http://www.ncbi.nlm.nih.gov/books/NBK569880/),ISBN 978\-92\-4\-000012\-4 978\-92\-4\-000296\-8,LCCN NBK569880Cited by:[§IV\-A](https://arxiv.org/html/2605.21563#S4.SS1.p1.5)\.
- \[19\]F\. Zhang, D\. Kreuter, Y\. Chen, S\. Dittmer, S\. Tull, T\. Shadbahr, M\. Schut, F\. Asselbergs, S\. Kar, S\. Sivapalaratnam, S\. Williams, M\. Koh, Y\. Henskens, B\. De Wit, U\. D’Alessandro, B\. Bah, O\. Secka, P\. Nachev, R\. Gupta, S\. Trompeter, N\. Boeckx, C\. Van Laer, G\. A\. Awandare, K\. Sarpong, L\. Amenga\-Etego, M\. Leers, M\. Huijskens, S\. McDermott, W\. H\. Ouwehand, J\. Rudd, C\. Schönlieb, N\. Gleadall, M\. Roberts, J\. Preller, J\. H\.F\. Rudd, J\. A\.D\. Aston, C\. Schönlieb, N\. Gleadall, and M\. Roberts\(2024\-06\)Recent methodological advances in federated learning for healthcare\.5\(6\),pp\. 101006\.External Links:ISSN 26663899,[Document](https://dx.doi.org/10.1016/j.patter.2024.101006)Cited by:[§I](https://arxiv.org/html/2605.21563#S1.p2.1)\.
- \[20\]F\. Zhang, D\. Kreuter, C\. Esteve\-Yagüe, S\. Dittmer, J\. Fernandez\-Marques, S\. Ip, B\. Consortium, N\. C\. J\. de Wit, A\. Wood, J\. H\. Rudd, N\. Lane, N\. S\. Gleadall, C\. Schönlieb, and M\. Roberts\(2025\)FedMAP: personalised federated learning for real large\-scale healthcare systems\.External Links:2405\.19000,[Link](https://arxiv.org/abs/2405.19000)Cited by:[§I](https://arxiv.org/html/2605.21563#S1.p3.1),[§II](https://arxiv.org/html/2605.21563#S2.p1.1),[3rd item](https://arxiv.org/html/2605.21563#S4.I1.i3.p1.1),[Code Availability](https://arxiv.org/html/2605.21563#Sx3.p1.1)\.
- \[21\]F\. Zhang, D\. Kreuter, J\. Fernandez\-Marques, B\. Consortium, G\. Verghese, B\. Butler, N\. Lane, S\. Sivapalaratnam, J\. Taylor, N\. C\. J\. de Wit, N\. S\. Gleadall, C\. Schönlieb, and M\. Roberts\(2026\)Building privacy\-and\-security\-focused federated learning infrastructure for global multi\-centre healthcare research\.arXiv preprint arXiv:2603\.10063\.Cited by:[§I](https://arxiv.org/html/2605.21563#S1.p2.1),[§I](https://arxiv.org/html/2605.21563#S1.p3.1),[§V](https://arxiv.org/html/2605.21563#S5.p1.1),[Code Availability](https://arxiv.org/html/2605.21563#Sx3.p1.1)\.